Standard OTel showed 7 failures — there were 18: what MCP tracing misses
airesearch12 · x · 2026-10-02
MCP tools typically return errors as isError: true inside a successful JSON-RPC response — the transport says 200 OK, so standard OpenTelemetry records a clean green span while the agent is actually receiving permissiondenied, notfound, or validationerror. The author's library opentel-mcp catches these; its new release ships a local dashboard runnable with one command (npx opentel-mcp-ui --demo, no Prometheus/Grafana/Collector needed) showing how many failures standard OTel misses (11 of 18 in demo data) with side-by-side comparisons.
The library provides spans for tools/call and tools/list aligned with OTel MCP semantic conventions, 12 metrics (silent failures, errors, duration, tokens, cost, agent thrash, schema drift), error fingerprinting/classification, support for MCP SDK v1/v2 over stdio and HTTP, and it never throws into your server. Limits: server-side only, tools but not resources/prompts, cost tracking needs tools to report token usage.
Related event: Standard OpenTelemetry Misses Silent MCP Tool Failures(2 posts)→
More from coding & agent
- CloudAnalyzer: browser-based LiDAR tools for SLAM map fixes and Lanelet2 editing in Rust+WASM — rsasaki0109 · 2026-10-02
- AI Coding Accelerates PS5 Emulation: SharpEmu Now Runs 10 Playable Titles — mark_k · 2026-10-02
- Muse adds a main agent role to help users figure out what to automate — op7418 · 2026-10-02
- shardr: a content-addressed model store with BitTorrent sync and OpenAI-compatible serving — Cyb3erDudu · 2026-10-02
- Cloudflare adds Analytics SQL binding to query analytics datasets from Workers — ritakozlov · 2026-10-02
- a16z podcast: Lio CEO on why AI-native startups win work outside the system of record — a16z · 2026-10-02